query result cache in oracle 11g

Caching has been a feature of Oracle for as long as most of us can remember. Over the many years and versions, Oracle has continually expanded its caching mechanisms. We are all familiar with the buffer cache, keep cache, library cache, shared pool, PGA/UGA and so on. In addition to the many data and cursor caches included in its architecture, Oracle has also enhanced caching support for common programming techniques; for example, scalar subquery caching, global temporary tables and associative arrays. In 11g, Oracle has extended this further by adding result caching to both its server and client architecture.

There are three new result caching features in 11g:

This article will describe and demonstrate the query result cache only. The PL/SQL function result cache feature shares much of the same architecture, but will be the subject of a future article.

an overview

As its name suggests, the query result cache is used to store the results of SQL queries for re-use in subsequent executions. By caching the results of queries, Oracle can avoid having to repeat the potentially time-consuming and intensive operations that generated the resultset in the first place (for example, sorting/aggregation, physical I/O, joins etc). The cache results themselves are available across the instance (i.e. for use by sessions other than the one that first executed the query) and are maintained by Oracle in a dedicated area of memory. Unlike our homegrown solutions using associative arrays or global temporary tables, the query result cache is completely transparent to our applications. It is also maintained for consistency automatically, unlike our own caching programs.

We will examine the features of the query result cache in more detail throughout this article.

database configuration

We will begin by looking at some of the database configuration required to use the query result cache. The initialisation parameters are as follows.

SQL> SELECT name, value, isdefault
  2  FROM   v$parameter
  3  WHERE  name LIKE 'result_cache%';

NAME                               VALUE              ISDEFAULT
---------------------------------- ------------------ ---------
result_cache_mode                  MANUAL             TRUE
result_cache_max_size              1081344            TRUE
result_cache_max_result            5                  TRUE
result_cache_remote_expiration     0                  TRUE

4 rows selected.

A brief explanation of each of these parameters is as follows.

The cache size is dynamic and can be changed either permanently or until the instance is restarted. We will roughly double the size of the cache for this article and verify that we have a larger result cache as follows (note this was run as SYSDBA).

SQL> ALTER SYSTEM SET result_cache_max_size = 2M SCOPE = MEMORY;

System altered.

SQL> SELECT name, value
  2  FROM   v$parameter
  3  WHERE  name = 'result_cache_max_size';

NAME                                     VALUE
---------------------------------------- -------------------------
result_cache_max_size                    2097152

1 row selected.

The setup for the result cache is simple and should be a one-time DBA operation. We will now see some examples of caching and using results below.

caching results manually

As we saw earlier, the default caching mode for this instance is MANUAL. This means that query resultsets will not be cached unless we instruct Oracle to do so by using the RESULT_CACHE hint. In our first example below, we will manually cache the results of a simple aggregate query. Note that the examples in this article are all based on the SH sample schema. First, we verify our cache mode as follows.

SQL> SELECT value
  2  FROM   v$parameter
  3  WHERE  name = 'result_cache_mode';

VALUE
----------------
MANUAL

1 row selected.

We will now run a query and cache its results. We will run this through Autotrace because we are interested in both the workload statistics and the execution plan (Autotrace will also conveniently suppress the query output).

SQL> set autotrace traceonly

SQL> set timing on

SQL> SELECT /*+ RESULT_CACHE */
  2         p.prod_name
  3  ,      SUM(s.amount_sold)   AS total_revenue
  4  ,      SUM(s.quantity_sold) AS total_sales
  5  FROM   sales s
  6  ,      products p
  7  WHERE  s.prod_id = p.prod_id
  8  GROUP  BY
  9         p.prod_name;

71 rows selected.

Elapsed: 00:00:05.00

Using the RESULT_CACHE hint, we have instructed Oracle to cache the results of this aggregate query. We can see that it returned 71 rows and took 5 seconds to execute. We will see the amount of work that Oracle did to generate these results further below, but first we will see the execution plan (note that this is a theoretical explain plan and not the real execution plan, but is a good approximation in this system).


Execution Plan
----------------------------------------------------------
Plan hash value: 504757596

----------------------------------------------------------------------- ... -----------------
| Id  | Operation                | Name                       | Rows  | ... | Pstart| Pstop |
----------------------------------------------------------------------- ... -----------------
|   0 | SELECT STATEMENT         |                            |    71 | ... |       |       |
|   1 |  RESULT CACHE            | 091zc7mvn8ums36mbd2gqac4h0 |       | ... |       |       |
|   2 |   HASH GROUP BY          |                            |    71 | ... |       |       |
|*  3 |    HASH JOIN             |                            |    72 | ... |       |       |
|   4 |     VIEW                 | VW_GBC_5                   |    72 | ... |       |       |
|   5 |      HASH GROUP BY       |                            |    72 | ... |       |       |
|   6 |       PARTITION RANGE ALL|                            |   918K| ... |     1 |    28 |
|   7 |        TABLE ACCESS FULL | SALES                      |   918K| ... |     1 |    28 |
|   8 |     TABLE ACCESS FULL    | PRODUCTS                   |    72 | ... |       |       |
----------------------------------------------------------------------- ... -----------------

Predicate Information (identified by operation id):
---------------------------------------------------

   3 - access("ITEM_1"="P"."PROD_ID")

Result Cache Information (identified by operation id):
------------------------------------------------------

   1 - column-count=3; dependencies=(SH.SALES, SH.PRODUCTS); parameters=(nls); name="SELECT /*+ RESULT_CACHE */
       p.prod_name
,      SUM(s.amount_sold)   AS total_revenue
,      SUM(s.quantity_sold) AS total_"

Note the highlighted sections of the execution plan. It contains some new information, which we can summarise as follows:

Finally, the Autotrace report displays the work that Oracle performed to generate these results.


Statistics
----------------------------------------------------------
      14871  recursive calls
          0  db block gets
       4890  consistent gets
       1745  physical reads
          0  redo size
       3526  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
        136  sorts (memory)
          0  sorts (disk)
         71  rows processed

We can see a range of I/O and CPU activity in these figures, as expected. We will now test the new query result cache by running the same query a second time and comparing the Autotrace report, as follows.

SQL> SELECT /*+ RESULT_CACHE */
  2         p.prod_name
  3  ,      SUM(s.amount_sold)   AS total_revenue
  4  ,      SUM(s.quantity_sold) AS total_sales
  5  FROM   sales s
  6  ,      products p
  7  WHERE  s.prod_id = p.prod_id
  8  GROUP  BY
  9         p.prod_name;

71 rows selected.

Elapsed: 00:00:00.01

Execution Plan
----------------------------------------------------------
Plan hash value: 504757596

----------------------------------------------------------------------- ... -----------------
| Id  | Operation                | Name                       | Rows  | ... | Pstart| Pstop |
----------------------------------------------------------------------- ... -----------------
|   0 | SELECT STATEMENT         |                            |    71 | ... |       |       |
|   1 |  RESULT CACHE            | 091zc7mvn8ums36mbd2gqac4h0 |       | ... |       |       |
|   2 |   HASH GROUP BY          |                            |    71 | ... |       |       |
|*  3 |    HASH JOIN             |                            |    72 | ... |       |       |
|   4 |     VIEW                 | VW_GBC_5                   |    72 | ... |       |       |
|   5 |      HASH GROUP BY       |                            |    72 | ... |       |       |
|   6 |       PARTITION RANGE ALL|                            |   918K| ... |     1 |    28 |
|   7 |        TABLE ACCESS FULL | SALES                      |   918K| ... |     1 |    28 |
|   8 |     TABLE ACCESS FULL    | PRODUCTS                   |    72 | ... |       |       |
----------------------------------------------------------------------- ... -----------------

Predicate Information (identified by operation id):
---------------------------------------------------

   3 - access("ITEM_1"="P"."PROD_ID")

Result Cache Information (identified by operation id):
------------------------------------------------------

   1 - column-count=3; dependencies=(SH.SALES, SH.PRODUCTS); parameters=(nls); name="SELECT /*+ RESULT_CACHE */
       p.prod_name
,      SUM(s.amount_sold)   AS total_revenue
,      SUM(s.quantity_sold) AS total_"


Statistics
----------------------------------------------------------
          0  recursive calls
          0  db block gets
          0  consistent gets
          0  physical reads
          0  redo size
       3526  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
         71  rows processed

Starting with the statistics report, we can see that this time Oracle has done very little work. In fact it has performed none of the I/O, sorting or recursive SQL that was required to answer our query the first time. Oracle has recognised that the query can be satisfied from the result cache and simply returned the pre-computed answer to us instead, in approximately 0.1 seconds.

Interestingly, the execution plan remains the same (this is to be expected because the SQL is not optimised a second time) but is now slightly misleading. None of the plan operations actually take place once we have a reusable resultset, but the presence of the RESULT CACHE operation should alert us to the fact that we might already have a cached set of results. In fact, we can use the information supplied in this plan to verify the existence of a cached resultset for ourselves, which we will examine later in this article.

We have now seen a simple example of query result caching. Minimising the amount of work that Oracle has to do to answer our query will reduce the time it takes. It also follows that the more work Oracle can avoid, the better the gains from caching.

automatic result caching

The alternative result_cache_mode to MANUAL is FORCE. This can be session or system specific and in this mode Oracle will attempt to set or use cached query results when it can, unless we use the NO_RESULT_CACHE hint. We will see an example of this mode below. We will set the mode to FORCE at a session level, then repeat our previous SQL example minus the RESULT_CACHE hint. First we set the result_cache_mode as follows.

SQL> ALTER SESSION SET result_cache_mode = FORCE;

Session altered.

With Autotrace set, we will now run the same aggregate statement from our manual example above. Note that the only modification to this SQL statement is to remove the RESULT_CACHE hint.

SQL> set autotrace traceonly

SQL> set timing on

SQL> SELECT p.prod_name
  2  ,      SUM(s.amount_sold)   AS total_revenue
  3  ,      SUM(s.quantity_sold) AS total_sales
  4  FROM   sales s
  5  ,      products p
  6  WHERE  s.prod_id = p.prod_id
  7  GROUP  BY
  8         p.prod_name;

71 rows selected.

Elapsed: 00:00:00.01

Execution Plan
----------------------------------------------------------
Plan hash value: 504757596

----------------------------------------------------------------------- ... -----------------
| Id  | Operation                | Name                       | Rows  | ... | Pstart| Pstop |
----------------------------------------------------------------------- ... -----------------
|   0 | SELECT STATEMENT         |                            |    71 | ... |       |       |
|   1 |  RESULT CACHE            | 091zc7mvn8ums36mbd2gqac4h0 |       | ... |       |       |
|   2 |   HASH GROUP BY          |                            |    71 | ... |       |       |
|*  3 |    HASH JOIN             |                            |    72 | ... |       |       |
|   4 |     VIEW                 | VW_GBC_5                   |    72 | ... |       |       |
|   5 |      HASH GROUP BY       |                            |    72 | ... |       |       |
|   6 |       PARTITION RANGE ALL|                            |   918K| ... |     1 |    28 |
|   7 |        TABLE ACCESS FULL | SALES                      |   918K| ... |     1 |    28 |
|   8 |     TABLE ACCESS FULL    | PRODUCTS                   |    72 | ... |       |       |
----------------------------------------------------------------------- ... -----------------

Predicate Information (identified by operation id):
---------------------------------------------------

   3 - access("ITEM_1"="P"."PROD_ID")

Result Cache Information (identified by operation id):
------------------------------------------------------

   1 - column-count=3; dependencies=(SH.SALES, SH.PRODUCTS); parameters=(nls); name="SELECT p.prod_name
,      SUM(s.amount_sold)   AS total_revenue
,      SUM(s.quantity_sold) AS total_sales
FROM   sales s
,     "

This is really interesting! We can see that in FORCE mode without a RESULT_CACHE hint, Oracle has used the same plan as the "equivalent" statement in MANUAL mode with the RESULT_CACHE hint (though in reality they are different cursors). Note that the result cache hashes are the same, although the name parameter is different. Given this, we would expect Oracle to use our previously cached results, which we can verify from the Statistics report, as follows.


Statistics
----------------------------------------------------------
          1  recursive calls
          0  db block gets
          0  consistent gets
          0  physical reads
          0  redo size
       3526  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
         71  rows processed

We can see quite clearly that Oracle has returned the results that it cached from the manual-mode, hinted SQL. It recognises their equivalence despite the difference in the SQL text.

Continuing in FORCE result_cache_mode, we will now execute a new aggregate statement, as follows.

SQL> SELECT p.prod_name
  2  ,      MAX(s.quantity_sold) AS max_sales
  3  FROM   sales s
  4  ,      products p
  5  WHERE  s.prod_id = p.prod_id
  6  GROUP  BY
  7         p.prod_name;

71 rows selected.

Elapsed: 00:00:00.59

Execution Plan
----------------------------------------------------------
Plan hash value: 504757596

----------------------------------------------------------------------- ... -----------------
| Id  | Operation                | Name                       | Rows  | ... | Pstart| Pstop |
----------------------------------------------------------------------- ... -----------------
|   0 | SELECT STATEMENT         |                            |    71 | ... |       |       |
|   1 |  RESULT CACHE            | 12scakxrxks3p73w5nxr69wn3j |       | ... |       |       |
|   2 |   HASH GROUP BY          |                            |    71 | ... |       |       |
|*  3 |    HASH JOIN             |                            |    72 | ... |       |       |
|   4 |     VIEW                 | VW_GBC_5                   |    72 | ... |       |       |
|   5 |      HASH GROUP BY       |                            |    72 | ... |       |       |
|   6 |       PARTITION RANGE ALL|                            |   918K| ... |     1 |    28 |
|   7 |        TABLE ACCESS FULL | SALES                      |   918K| ... |     1 |    28 |
|   8 |     TABLE ACCESS FULL    | PRODUCTS                   |    72 | ... |       |       |
----------------------------------------------------------------------- ... -----------------

Predicate Information (identified by operation id):
---------------------------------------------------

   3 - access("ITEM_1"="P"."PROD_ID")

Result Cache Information (identified by operation id):
------------------------------------------------------

   1 - column-count=2; dependencies=(SH.SALES, SH.PRODUCTS); parameters=(nls); name="SELECT p.prod_name
,      MAX(s.quantity_sold) AS max_sales
FROM   sales s
,      products p
WHERE  s.prod_id = p.prod_id
GROUP "


Statistics
----------------------------------------------------------
          1  recursive calls
          0  db block gets
       1731  consistent gets
          0  physical reads
          0  redo size
       2687  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
         71  rows processed

In FORCE mode, Oracle has cached the results of our query without us requesting it. We can see this in the execution plan. We will run the query a second time, as follows.

SQL> SELECT p.prod_name
  2  ,      MAX(s.quantity_sold) AS max_sales
  3  FROM   sales s
  4  ,      products p
  5  WHERE  s.prod_id = p.prod_id
  6  GROUP  BY
  7         p.prod_name;

71 rows selected.

Elapsed: 00:00:00.00

<< ...plan removed... >>

Statistics
----------------------------------------------------------
          0  recursive calls
          0  db block gets
          0  consistent gets
          0  physical reads
          0  redo size
       2687  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
         71  rows processed

The wall-clock timing and Statistics report shows clearly that Oracle has retrieved the cached results.

dynamic result cache views

So far we have seen the effects of caching with the two modes of the query result cache. We will now look a little deeper into what happens with the query cache and what information Oracle exposes about it. We can search the data dictionary for the result cache dynamic views, as follows.

SQL> SELECT view_name
  2  FROM   dba_views
  3  WHERE  view_name LIKE 'V_$RESULT_CACHE%';

VIEW_NAME
------------------------------
V_$RESULT_CACHE_DEPENDENCY
V_$RESULT_CACHE_MEMORY
V_$RESULT_CACHE_OBJECTS
V_$RESULT_CACHE_STATISTICS

4 rows selected.

Oracle provides four dynamic views. We will have a brief look at these below (refer to the online documentation for more details: a link is provided at the end of this article). We will start with V$RESULT_CACHE_OBJECTS, which exposes the most information about our cached query results.

SQL> DESC v$result_cache_objects;

           Name                     Null?  Type
           ------------------------ ------ ------------------
    1      ID                              NUMBER
    2      TYPE                            VARCHAR2(10)
    3      STATUS                          VARCHAR2(9)
    4      BUCKET_NO                       NUMBER
    5      HASH                            NUMBER
    6      NAME                            VARCHAR2(128)
    7      NAMESPACE                       VARCHAR2(5)
    8      CREATION_TIMESTAMP              DATE
    9      CREATOR_UID                     NUMBER
   10      DEPEND_COUNT                    NUMBER
   11      BLOCK_COUNT                     NUMBER
   12      SCN                             NUMBER
   13      COLUMN_COUNT                    NUMBER
   14      PIN_COUNT                       NUMBER
   15      SCAN_COUNT                      NUMBER
   16      ROW_COUNT                       NUMBER
   17      ROW_SIZE_MAX                    NUMBER
   18      ROW_SIZE_MIN                    NUMBER
   19      ROW_SIZE_AVG                    NUMBER
   20      BUILD_TIME                      NUMBER
   21      LRU_NUMBER                      NUMBER
   22      OBJECT_NO                       NUMBER
   23      INVALIDATIONS                   NUMBER
   24      SPACE_OVERHEAD                  NUMBER
   25      SPACE_UNUSED                    NUMBER
   26      CACHE_ID                        VARCHAR2(93)
   27      CACHE_KEY                       VARCHAR2(93)

As we can see, there is a great deal of information potentially available. We will query a few important columns below.

SQL> SELECT name
  2  ,      type
  3  ,      cache_id
  4  ,      row_count
  5  FROM   v$result_cache_objects
  6  ORDER  BY
  7         creation_timestamp;

NAME                           TYPE       CACHE_ID                    ROW_COUNT
------------------------------ ---------- -------------------------- ----------
SH.PRODUCTS                    Dependency SH.PRODUCTS                         0
SH.SALES                       Dependency SH.SALES                            0
SELECT /*+ RESULT_CACHE */     Result     091zc7mvn8ums36mbd2gqac4h0         71
SELECT p.prod_name             Result     12scakxrxks3p73w5nxr69wn3j         71
SELECT DECODE('A','A','1','2'  Result     0y8dgk314f9f8bz05qsrrny8u8          1

5 rows selected.

We can see two types of information in this view: dependencies and results. We will discuss dependencies later, but the results' names clearly align with the queries we have run so far (the SUM and MAX aggregate sales queries). The last query in the output is executed by SQL*Plus. Remember from earlier that we executed two SQL statements (equivalent except for the RESULT_CACHE hint) and note the CACHE_ID values. There is only one entry for the two statements due to the fact that they shared a result set and hashed to the same CACHE_ID.

We can also look at the result cache statistics for a high-level overview of how it is being used, as follows.

SQL> SELECT *
  2  FROM   v$result_cache_statistics;

        ID NAME                                        VALUE
---------- --------------------------------- ---------------
         1 Block Size (Bytes)                           1024
         2 Block Count Maximum                          2048
         3 Block Count Current                            32
         4 Result Size Maximum (Blocks)                  102
         5 Create Count Success                            3
         6 Create Count Failure                            0
         7 Find Count                                      4
         8 Invalidation Count                              0
         9 Delete Count Invalid                            0
        10 Delete Count Valid                              0

10 rows selected.

We can see some general information on cache entries and hits, which are minimal so far as we have used a clean cache to run our examples. We can also see some statistics about how our cache is organised (i.e. up to 2,048 blocks of 1,024 bytes each which is equivalent to the 2M result_cache_max_size value that we set earlier). Currently, we are using 32K of the cache (32 blocks) and no single result set can be more than 102 blocks (or consume 102K of memory).

result cache dependencies

Each query result is dependant on one or more tables (i.e. the source tables for the query). We can get information on which objects a query is dependant on in a number of places. The V$RESULT_CACHE_DEPENDENCY view summarises the dependencies for each entry in the result cache. We saw the dependencies parameter in the Result Cache report from DBMS_XPLAN.DISPLAY which listed the tables involved in our sample aggregate queries. We also saw entries in the V$RESULT_CACHE_OBJECTS view data with a type of "Dependency". We can put these together to summarise the dependencies as follows.

SQL> SELECT ro.id
  2  ,      ro.name
  3  ,      wm_concat(do.object_name) AS object_names
  4  FROM   v$result_cache_objects    ro
  5            LEFT OUTER JOIN
  6         v$result_cache_dependency rd
  7            ON (ro.id = rd.result_id)
  8            LEFT OUTER JOIN
  9         dba_objects               do
 10            ON (rd.object_no = do.object_id)
 11  WHERE  ro.type = 'Result'
 12  GROUP  BY
 13         ro.id
 14  ,      ro.name;

        ID NAME                                               OBJECT_NAMES
---------- -------------------------------------------------- ----------------
         2 SELECT /*+ RESULT_CACHE */                         SALES,PRODUCTS
                  p.prod_name
           ,      SUM(s.amount_sold)   AS total_revenue
           ,      SUM(s.quantity_sold) AS total_

         6 SELECT DECODE('A','A','1','2') FROM DUAL
         7 SELECT p.prod_name                                 SALES,PRODUCTS
           ,      MAX(s.quantity_sold) AS max_sales
           FROM   sales s
           ,      products p
           WHERE  s.prod_id = p.prod_id
           GROUP


3 rows selected.

Dependencies are necessary to protect the integrity of the query results in the cache. If the data in any of the dependant tables is modified, Oracle will invalidate the result cache entry and will not use it until it is refreshed by a repeat of the original SQL. This behaviour cannot be circumvented, even if we are prepared to tolerate inconsistent results.

We can demonstrate result cache invalidation very easily. We will perform a "no-change" update to a single row of PRODUCTS and commit the transaction, as follows.

SQL> UPDATE products
  2  SET    prod_name = prod_name
  3  WHERE  ROWNUM = 1;

1 row updated.

SQL> COMMIT;

Commit complete.

We will now repeat one of our cached aggregation queries and measure the workload using Autotrace.

SQL> set autotrace traceonly statistics

SQL> SELECT p.prod_name
  2  ,      MAX(s.quantity_sold) AS max_sales
  3  FROM   sales s
  4  ,      products p
  5  WHERE  s.prod_id = p.prod_id
  6  GROUP  BY
  7         p.prod_name;

71 rows selected.

Statistics
----------------------------------------------------------
          0  recursive calls
          0  db block gets
       1731  consistent gets
          0  physical reads
          0  redo size
       2687  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
         71  rows processed

Oracle will not attempt to understand the nature of the modification to the dependant objects. Even with a no-change update, the cached result entry was invalidated and the subsequent repeat of the source SQL caused the data to be generated again. The V$RESULT_CACHE_OBJECTS view provides some statistics on this, as follows.

SQL> SELECT id
  2  ,      name
  3  ,      type
  4  ,      invalidations
  5  ,      status
  6  FROM   v$result_cache_objects
  7  ORDER  BY
  8         id;

  ID NAME                            TYPE       INVALIDATIONS STATUS
---- ------------------------------- ---------- ------------- ---------
   1 SH.PRODUCTS                     Dependency             1 Published
   0 SH.SALES                        Dependency             0 Published
   6 SELECT DECODE('A','A','1','2')  Result                 0 Published
   2 SELECT /*+ RESULT_CACHE */      Result                 0 Invalid
   7 SELECT p.prod_name              Result                 0 Invalid
  10 SELECT p.prod_name              Result                 0 Published

6 rows selected.

We can see that the invalidation occurred at two levels. First, the INVALIDATIONS column details the number of times that modifications to an underlying table have caused an invalidation. Second, the STATUS column shows us which results have been invalidated by the same action. When we updated the PRODUCTS table, we invalidated the results from our previous queries (IDs 2 and 7). We then repeated one of the original queries, for which Oracle created a new set of results in the cache (ID 10).

cache find count

If we are caching query results, we might be interested to know how often they are used. The V$RESULT_CACHE_STATISTICS view provides a "Find Count" statistic, but this is cache-wide so we can't limit it to a particular query. In the following example, we will capture the current Find Count and then run a SQL statement in a PL/SQL loop 100 times.

SQL> SELECT value
  2  FROM   v$result_cache_statistics
  3  WHERE  name = 'Find Count';

          VALUE
---------------
              6

1 row selected.

SQL> DECLARE
  2     n PLS_INTEGER;
  3  BEGIN
  4     FOR i IN 1 .. 100 LOOP
  5        SELECT /*+ RESULT_CACHE */ COUNT(*) INTO n FROM channels;
  6     END LOOP;
  7  END;
  8  /

PL/SQL procedure successfully completed.

We will now measure the Find Count again, as follows.

SQL> SELECT value
  2  FROM   v$result_cache_statistics
  3  WHERE  name = 'Find Count';

          VALUE
---------------
            105

1 row selected.

This increased by 99, which is to be expected. We executed our SQL statement 100 times. The first execution cached the results and the 99 remaining executions used them. Needless to say, this was a single-user test system. We can confirm that we added the SQL results to the cache as follows.

SQL> SELECT name
  2  ,      type
  3  ,      row_count
  4  FROM   v$result_cache_objects
  5  ORDER  BY
  6         creation_timestamp;

NAME                                               TYPE        ROW_COUNT
-------------------------------------------------- ---------- ----------
SELECT /*+ RESULT_CACHE */                         Result             71
SH.PRODUCTS                                        Dependency          0
SH.SALES                                           Dependency          0
SELECT p.prod_name                                 Result             71
SELECT DECODE('A','A','1','2') FROM DUAL           Result              1
SELECT p.prod_name                                 Result             71
SH.CHANNELS                                        Dependency          0
SELECT /*+ RESULT_CACHE */ COUNT(*) FROM CHANNELS  Result              1

8 rows selected.

Hence there is no difference for SQL embedded in PL/SQL programs.

parameterised caching

In our previous example, we executed a single SQL statement 100 times and saw 1 cache entry. A far more common scenario is to have single-row lookups based on a primary key derived from another cursor (this is not particularly efficient, but is still extremely common). The query result cache handles this scenario by recognising the different bind variables and caching each resultset independently. The bind variables act as parameters to the result cache lookup and are listed in the Result Cache report from DBMS_XPLAN. If a bind variable is repeated, the cached results will be used.

We will demonstrate this behaviour below. We will set the result_cache_mode to FORCE for convenience. We will choose 4 products and lookup each one 10 times. The lookup will use bind variables.

SQL> ALTER SESSION SET result_cache_mode = FORCE;

Session altered.

SQL> DECLARE
  2  
  3     TYPE id_ntt IS TABLE OF products.prod_id%TYPE;
  4     nt_ids id_ntt := id_ntt(40,41,42,43);
  5  
  6     v_name products.prod_name%TYPE;
  7  
  8  BEGIN
  9     FOR i IN 1 .. 10 LOOP
 10        FOR ii IN 1 .. nt_ids.COUNT LOOP
 11  
 12           SELECT prod_name INTO v_name
 13           FROM   products
 14           WHERE  prod_id = nt_ids(ii);
 15  
 16        END LOOP;
 17     END LOOP;
 18  END;
 19  /

PL/SQL procedure successfully completed.

According to what we now know about the result cache mechanism, we ran 4 different SQL statements above (the same SQL statement with 4 different inputs). We will query V$RESULT_CACHE_OBJECTS to verify this, as follows.

SQL> SELECT name
  2  ,      type
  3  ,      row_count
  4  FROM   v$result_cache_objects
  5  ORDER  BY
  6         creation_timestamp;

NAME                                                    TYPE        ROW_COUNT
------------------------------------------------------- ---------- ----------
SELECT /*+ RESULT_CACHE */                              Result             71
<< ...snip... >>
SH.CHANNELS                                             Dependency          0
SELECT /*+ RESULT_CACHE */ COUNT(*) FROM CHANNELS       Result              1
SELECT PROD_NAME FROM PRODUCTS WHERE PROD_ID = :B1      Result              1
SELECT PROD_NAME FROM PRODUCTS WHERE PROD_ID = :B1      Result              1
SELECT PROD_NAME FROM PRODUCTS WHERE PROD_ID = :B1      Result              1
SELECT PROD_NAME FROM PRODUCTS WHERE PROD_ID = :B1      Result              1

12 rows selected.

We can see that the results for the same SQL text was added to the cache 4 times, as expected. The bind variable inputs are additional parameters to the cache lookup. Each resultset was added on the first execution of each cursor and the cache was "hit" 9 times for each cursor. Similar logic is commonly used by developers in associative array caching; a colleague of mine calls this "on-demand caching" (i.e. rather than cache entire lookup tables, only cache a lookup record when it is actually requested).

Remember that the result_cache_max_result parameter specifies that the largest cached resultset possible is n% of the total cache memory. While this protects us from filling the cache with the results of a single SQL statement, it doesn't stop us from filling the cache with parameterised cursors like those we saw above. In the following example, we will lookup every customer in the CUSTOMERS table twice. Again, we will be in FORCE result_cache_mode for convenience. Note that there are 55,500 records in the SH.CUSTOMERS demo table.

SQL> ALTER SESSION SET result_cache_mode = FORCE;

Session altered.

SQL> DECLARE
  2     v_first_name customers.cust_first_name%TYPE;
  3  BEGIN
  4     FOR i IN 1 .. 2 LOOP
  5        FOR r IN (SELECT cust_id FROM customers) LOOP
  6           SELECT cust_first_name INTO v_first_name
  7           FROM   customers
  8           WHERE  cust_id = r.cust_id;
  9        END LOOP;
 10     END LOOP;
 11  END;
 12  /

PL/SQL procedure successfully completed.

We will examine the cache entries below. Based on what we know about result cache behaviour, we can expect a large number of single-row resultsets, so we will try to aggregate these. We will query the minimum and maximum names in V$RESULT_CACHE_OBJECTS, together with a count of the entries, as follows.

SQL> SELECT MIN(name) AS min_name
  2  ,      MAX(name) AS max_name
  3  ,      COUNT(*)  AS cache_entries
  4  FROM   v$result_cache_objects
  5  WHERE  type = 'Result';

MIN_NAME                       MAX_NAME                       CACHE_ENTRIES
------------------------------ ------------------------------ -------------
SELECT CUST_FIRST_NAME FROM CU SELECT CUST_FIRST_NAME FROM CU          2035
STOMERS WHERE CUST_ID = :B1    STOMERS WHERE CUST_ID = :B1

1 row selected.

We added 2,035 customer lookups (out of a possible 55,500) to the cache. In fact, we completely flushed our previous results from the cache. We should therefore be aware of the potential for single lookups, particularly in PL/SQL programs, to "hog" the cache. If we query V$RESULT_CACHE_STATISTICS, we will see that the "Create Count Success" statistic should be quite high.

SQL> SELECT *
  2  FROM   v$result_cache_statistics;

        ID NAME                                      VALUE
---------- ------------------------------- ---------------
         1 Block Size (Bytes)                         1024
         2 Block Count Maximum                        2048
         3 Block Count Current                        2048
         4 Result Size Maximum (Blocks)                102
         5 Create Count Success                     111011
         6 Create Count Failure                          0
         7 Find Count                                  141
         8 Invalidation Count                            2
         9 Delete Count Invalid                          4
        10 Delete Count Valid                       108972

10 rows selected.

We have added over 111,000 resultsets to the cache, mostly as a result of the previous example. The loop through 55,500 customers would have continually replaced the existing cache entries (we only had room in the cache for approximately 4% of the total resultsets being processed in the PL/SQL).

flashback query results

The query result cache supports flashback queries. Most readers will be aware of flashback queries (an overview is available here). An SCN or timestamp is supplied to a flashback query using the AS OF extension to the table(s) in the FROM clause. This supplied point-in-time is treated by Oracle as a parameter to the query result cache.

To demonstrate this, we will run a simple flashback query twice. We will use Autotrace to demonstrate the result cache behaviour. We will begin by setting the result_cache_mode to FORCE for convenience.

SQL> ALTER SESSION SET result_cache_mode = FORCE;

Session altered.

We will setup a bind variable for our timestamp and execute a simple flashback query, as follows.

SQL> exec :ts := TO_CHAR(TRUNC(SYSDATE,'HH'),'YYYYMMDDHH24MISS');

PL/SQL procedure successfully completed.

SQL> set autotrace traceonly

SQL> SELECT MIN(prod_id)
  2  FROM   products AS OF TIMESTAMP TO_TIMESTAMP(:ts,'YYYYMMDDHH24MISS');

1 row selected.


Execution Plan
----------------------------------------------------------
Plan hash value: 1489483397

--------------------------------------------------------------------------- ... --
| Id  | Operation                   | Name                       | Rows  |  ...  |
--------------------------------------------------------------------------- ... --
|   0 | SELECT STATEMENT            |                            |     1 |  ...  |
|   1 |  RESULT CACHE               | 4vff36vw5vmn32gftq4a5qfpxh |       |  ...  |
|   2 |   SORT AGGREGATE            |                            |     1 |  ...  |
|   3 |    INDEX FULL SCAN (MIN/MAX)| PRODUCTS_PK                |    72 |  ...  |
--------------------------------------------------------------------------- ... --

Result Cache Information (identified by operation id):
------------------------------------------------------

   1 - column-count=1; attributes=(single-row); parameters=(:TS); name="SELECT MIN(prod_id)
FROM   products AS OF TIMESTAMP TO_TIMESTAMP(:ts,'YYYYMMDDHH24MISS')"


Statistics
----------------------------------------------------------
          3  recursive calls
          0  db block gets
         73  consistent gets
          0  physical reads
          0  redo size
        422  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
          1  rows processed

The Result Cache Information report provided by DBMS_XPLAN includes the parameters that Oracle used for this query; in this case the TS bind variable. The execution plan confirms that Oracle will cache the results of this flashback query. We will run the same query a second time to see if the results are re-used, as follows.

SQL> SELECT MIN(prod_id)
  2  FROM   products AS OF TIMESTAMP TO_TIMESTAMP(:ts,'YYYYMMDDHH24MISS');

1 row selected.

<< ...plan removed... >>

Statistics
----------------------------------------------------------
          1  recursive calls
          0  db block gets
          0  consistent gets
          0  physical reads
          0  redo size
        422  bytes sent via SQL*Net to client
        416  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
          1  rows processed

Given the same inputs (i.e. the same bind variable), Oracle was able to retrieve the flashback query results from the result cache, as confirmed by the Autotrace Statistics report.

performance considerations

Recall our examples so far. First, we ran some aggregate "report-style" queries to demonstrate the caching mechanism. These are probably not the types of queries we would expect to cache and re-use (they are primarily in the style of run-once reports). Second, we ran some lookup examples from PL/SQL, both with and without primary key predicates. These are far more likely to be candidates for caching in our applications: single-row lookups are common in PL/SQL programs. For this reason, we will concentrate on lookups to test the performance of the query result cache.

In the following example, we will use a version of Tom Kyte's Runstats program to compare the performance of 1 million cached and uncached single-row lookups. First, we will switch to MANUAL result_cache_mode, as follows.

SQL> ALTER SESSION SET result_cache_mode = MANUAL;

Session altered.

We will select 8 primary key values from the SH.PRODUCTS table and perform 125,000 lookups for each (totalling 1 million executions). We will begin with the uncached version, as follows.

SQL> exec runstats_pkg.rs_start;

PL/SQL procedure successfully completed.

SQL> DECLARE
  2  
  3     TYPE id_ntt IS TABLE OF products.prod_id%TYPE;
  4     nt_ids id_ntt := id_ntt(40,41,42,43,44,45,46,47);
  5  
  6     v_name products.prod_name%TYPE;
  7  
  8  BEGIN
  9     FOR i IN 1 .. 125000 LOOP
 10        FOR ii IN 1 .. nt_ids.COUNT LOOP
 11  
 12           SELECT prod_name INTO v_name
 13           FROM   products
 14           WHERE  prod_id = nt_ids(ii);
 15  
 16        END LOOP;
 17     END LOOP;
 18  END;
 19  /

PL/SQL procedure successfully completed.

We will now repeat the example but cache the eight different cursors, as follows.

SQL> exec runstats_pkg.rs_middle;

PL/SQL procedure successfully completed.

SQL> DECLARE
  2  
  3     TYPE id_ntt IS TABLE OF products.prod_id%TYPE;
  4     nt_ids id_ntt := id_ntt(40,41,42,43,44,45,46,47);
  5  
  6     v_name products.prod_name%TYPE;
  7  
  8  BEGIN
  9     FOR i IN 1 .. 125000 LOOP
 10        FOR ii IN 1 .. nt_ids.COUNT LOOP
 11  
 12           SELECT /*+ RESULT_CACHE */ prod_name INTO v_name
 13           FROM   products
 14           WHERE  prod_id = nt_ids(ii);
 15  
 16        END LOOP;
 17     END LOOP;
 18  END;
 19  /

PL/SQL procedure successfully completed.

We will now report the major differences between the two approaches, as follows.

SQL> exec runstats_pkg.rs_stop(1000);
Run1 ran in 4184 hsecs
Run2 ran in 4059 hsecs
Run1 ran in 103.08% of the time


Name                                  Run1        Run2        Diff
STAT..session pga memory          -327,680      65,536     393,216
STAT..calls to kcmgrs            1,000,000           8    -999,992
STAT..index fetch by key         1,000,002          10    -999,992
STAT..rows fetched via callbac   1,000,002          10    -999,992
STAT..table fetch by rowid       1,000,002          10    -999,992
LATCH.shared pool simulator             97   1,000,090     999,993
LATCH.cache buffers chains       2,000,007          23  -1,999,984
STAT..consistent gets            2,000,006          22  -1,999,984
STAT..consistent gets - examin   2,000,006          22  -1,999,984
STAT..consistent gets from cac   2,000,006          22  -1,999,984
STAT..session logical reads      2,000,006          22  -1,999,984
LATCH.Result Cache: Latch                0   2,000,017   2,000,017
STAT..buffer is not pinned cou   3,000,004          28  -2,999,976


Run1 latches total versus run2 -- difference and pct
Run1        Run2        Diff        Pct
3,003,103   4,003,090     999,987     75.02%

PL/SQL procedure successfully completed.

We can see that there is little performance difference between the cached and uncached lookups. In fact, the result cache uses more latches (note the two million hits on the new Result Cache latch), although it is marginally quicker overall. These results are probably to be expected when we consider the work to be done in a single primary key lookup. First, the index and table data will be in the buffer cache and second, the amount of I/O saved by a lookup in the query result cache is minimal.

The best performance gains will be achieved when the result cache is preventing a large amount of database work from being repeated (such as in our aggregate queries earlier). Unfortunately, these types of queries are going to be rare in the result cache as the base tables will probably be frequently updated.

encapsulating cached results

Caching lookup tables (in KEEP caches, user-defined arrays etc) is nothing new in Oracle. The query result cache provides another mechanism for doing this. In the following example, we will encapsulate the PRODUCTS table in a view. We will include the RESULT_CACHE hint and then load the contents of PRODUCTS into the query result cache, as follows.

SQL> CREATE VIEW v_products
  2  AS
  3     SELECT /*+ RESULT_CACHE */ *
  4     FROM   products;

View created.

SQL> SELECT * FROM v_products;

<< ...snip... >>

72 rows selected.

We can use the cached results of PRODUCTS as follows. Note that we are gathering plan statistics to investigate whether this technique leads to an I/O-free join to PRODUCTS.

SQL> SELECT /*+ GATHER_PLAN_STATISTICS */
  2         p.prod_name
  3  ,      AVG(s.quantity_sold) AS avg_sales
  4  FROM   sales s
  5  ,     (SELECT * FROM v_products) p
  6  WHERE  s.prod_id = p.prod_id
  7  GROUP  BY
  8         p.prod_name;

<< ...snip... >>

71 rows selected.

We will query the DBMS_XPLAN.DISPLAY_CURSOR function to view the rowsource plan for our example query, as follows.

SQL> SELECT *
  2  FROM   TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL,NULL,'RUNSTATS_LAST'));
PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------------------
SQL_ID  33wh37bxzhf26, child number 0
-------------------------------------
SELECT /*+ GATHER_PLAN_STATISTICS */        p.prod_name ,
AVG(s.quantity_sold) AS avg_sales FROM   sales s ,     (SELECT * FROM
v_products) p WHERE  s.prod_id = p.prod_id GROUP  BY        p.prod_name

Plan hash value: 1128805752

----------------------------------------------------------------------------------------- ... -----------
| Id  | Operation               | Name                       | Starts | E-Rows | A-Rows | ... | Buffers |
----------------------------------------------------------------------------------------- ... -----------
|   1 |  HASH GROUP BY          |                            |      1 |     71 |     71 | ... |    1727 |
|*  2 |   HASH JOIN             |                            |      1 |     72 |     72 | ... |    1727 |
|   3 |    VIEW                 | VW_GBC_5                   |      1 |     72 |     72 | ... |    1727 |
|   4 |     HASH GROUP BY       |                            |      1 |     72 |     72 | ... |    1727 |
|   5 |      PARTITION RANGE ALL|                            |      1 |    918K|    918K| ... |    1727 |
|   6 |       TABLE ACCESS FULL | SALES                      |     28 |    918K|    918K| ... |    1727 |
|   7 |    VIEW                 | V_PRODUCTS                 |      1 |     72 |     72 | ... |       0 |
|   8 |     RESULT CACHE        | dm772rv5jrfhpf7019qp8srvzu |      1 |        |     72 | ... |       0 |
|   9 |      TABLE ACCESS FULL  | PRODUCTS                   |      0 |     72 |      0 | ... |       0 |
----------------------------------------------------------------------------------------- ... -----------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access("ITEM_1"="V_PRODUCTS"."PROD_ID")

Result Cache Information (identified by operation id):
------------------------------------------------------

   8 -


32 rows selected.

For some reason, we are missing the information from the Result Cache section, but we can see clearly on operation 9 that we incurred no I/O from the PRODUCTS table. We will see if the same is true of a join to the V_PRODUCTS view itself below.

SQL> SELECT /*+ GATHER_PLAN_STATISTICS */
  2         p.prod_name
  3  ,      AVG(s.quantity_sold) AS avg_sales
  4  FROM   sales s
  5  ,      v_products p
  6  WHERE  s.prod_id = p.prod_id
  7  GROUP  BY
  8         p.prod_name;

<< ...snip... >>

71 rows selected.

SQL> SELECT *
  2  FROM   TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL,NULL,'RUNSTATS_LAST'));
PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------------------
SQL_ID  78nw9vr6vwg7n, child number 0
-------------------------------------
SELECT /*+ GATHER_PLAN_STATISTICS */        p.prod_name ,
AVG(s.quantity_sold) AS avg_sales FROM   sales s ,      v_products p
WHERE  s.prod_id = p.prod_id GROUP  BY        p.prod_name

Plan hash value: 1128805752

----------------------------------------------------------------------------------------- ... -----------
| Id  | Operation               | Name                       | Starts | E-Rows | A-Rows | ... | Buffers |
----------------------------------------------------------------------------------------- ... -----------
|   1 |  HASH GROUP BY          |                            |      1 |     71 |     71 | ... |    1731 |
|*  2 |   HASH JOIN             |                            |      1 |     72 |     72 | ... |    1731 |
|   3 |    VIEW                 | VW_GBC_5                   |      1 |     72 |     72 | ... |    1727 |
|   4 |     HASH GROUP BY       |                            |      1 |     72 |     72 | ... |    1727 |
|   5 |      PARTITION RANGE ALL|                            |      1 |    918K|    918K| ... |    1727 |
|   6 |       TABLE ACCESS FULL | SALES                      |     28 |    918K|    918K| ... |    1727 |
|   7 |    VIEW                 | V_PRODUCTS                 |      1 |     72 |     72 | ... |       4 |
|   8 |     RESULT CACHE        | 1b3305bb7rs5sdmdydr6u8sb49 |      1 |        |     72 | ... |       4 |
|   9 |      TABLE ACCESS FULL  | PRODUCTS                   |      1 |     72 |     72 | ... |       4 |
----------------------------------------------------------------------------------------- ... -----------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access("ITEM_1"="P"."PROD_ID")

Result Cache Information (identified by operation id):
------------------------------------------------------

   8 -


32 rows selected.

This time, we have incurred some I/O against the PRODUCTS table. Therefore, if we wish to cache a frequently-used table, we will need to repeat whatever query cached the data in the first place.

dbms_result_cache

Finally, we will look at a new built-in package, DBMS_RESULT_CACHE. This package is primarily for DBAs and has a number of APIs to manage and report on the result cache. We will look at one API in particular; the FLUSH function and procedure. Oracle supplies this API in both formats and as its name suggests, it enables us to clear the result cache and its statistics. We can optionally retain either the free memory in the cache or the cache statistics, but the default is to flush everything, as we will see below (note this is executed as SYSDBA).

SQL> BEGIN
  2     DBMS_RESULT_CACHE.FLUSH;
  3  END;
  4  /

PL/SQL procedure successfully completed.

SQL> SELECT * FROM v$result_cache_objects;

no rows selected

SQL> SELECT * FROM v$result_cache_statistics;

        ID NAME                                                         VALUE
---------- -------------------------------------------------- ---------------
         1 Block Size (Bytes)                                            1024
         2 Block Count Maximum                                           2048
         3 Block Count Current                                              0
         4 Result Size Maximum (Blocks)                                   102
         5 Create Count Success                                             0
         6 Create Count Failure                                             0
         7 Find Count                                                       0
         8 Invalidation Count                                               0
         9 Delete Count Invalid                                             0
        10 Delete Count Valid                                               0

10 rows selected.

All of our query results have been removed and the statistics have been reset. Needless to say, flushing the cache is not something that we would expect to do very often, if at all.

As stated, there are several other APIs for result cache management. The online documentation has full details and a link is provided at the end of this article.

summary

In this article, we have taken a brief look at the query result cache and how it can be used to optimise commonly-used queries. We have seen that the cache can provide good performance gains when the amount of I/O and CPU work required to repeat the results is significant. Conversely, we have seen that the gains from low-intensity queries (such as primary key lookups) are marginal. Unfortunately, it is probably the case that lookup and reference queries are the ones that our applications will repeat frequently; in which case the query result cache will not necessarily provide the gains that we might imagine from such a feature.

further reading

For more information on the query result cache, read the Performance Tuning Guide. The Pythian blog has a range of articles on result cache scalability, available in three parts (Part 1, Part 2 and Part 3). The PL/SQL Packages and Types Reference has full details on the DBMS_RESULT_CACHE package and descriptions of the dynamic performance views can be found in the Database Reference. Finally, the version of Runstats used in this article can be downloaded here.

source code

The source code for the examples in this article can be downloaded from here.

Adrian Billington, December 2007

Back to Top